Intelligent Spectrum Sensing Method, System, Device and Medium under Non-Gaussian Interference in Cognitive Radio Networks

Abstract

An intelligent spectrum sensing method, system, device and medium under non-Gaussian interference in a cognitive radio network. The method includes: processing the observed signal using a non-linear pre-processor to construct a generalized covariance matrix and a generalized time-frequency covariance matrix, and splicing and combining the two into a joint covariance matrix; using a Transformer network to extract features from the constructed joint covariance matrix to construct a feature vector based on the network output; constructing a detection statistic and a detection threshold using the feature vectors under different hypothesis tests, and realizing spectrum sensing under non-Gaussian interference by comparing the detection statistic and the detection threshold; the system, device and medium are used for intelligent spectrum sensing under non-Gaussian interference in a cognitive radio network; the sensing effect of the present invention is better, providing technical support for cognitive wireless networks and effectively improving the anti-interference ability of intelligent wireless systems.

Description

Technical Field

[0001] The present invention belongs to the technical field of spectrum sensing in cognitive radio networks, and discloses an intelligent spectrum sensing method, system, device and medium under non-Gaussian interference in a cognitive radio network. Background Art

[0002] A cognitive radio network can dynamically utilize idle spectrum without occupying exclusive spectrum resources, thereby improving spectrum utilization. To prevent interference to primary users, a cognitive radio network needs to accurately and real-time achieve high-precision spectrum sensing. As one of the key technologies of a cognitive radio network, spectrum sensing technology can sense the spectrum state in real time, so as to guide the cognitive radio network to achieve intelligent opportunistic transmission. Therefore, in a cognitive radio network, in order to realize the coexistence of primary user devices and secondary users, it is necessary to study efficient and accurate spectrum sensing methods.

[0003] Regarding the spectrum sensing problem, a variety of methods have been proposed. Traditional spectrum sensing methods can be roughly divided into: matched filtering algorithms (Chen H, Gao W and Daut D G. Signature based spectrum sensing algorithms for IEEE 802.22 WRAN[C]. IEEE International Conference on Communications, Glasgow, UK. 2007:6487-6492.), energy detection algorithms (Urkowitz H. Energy detection of unknown deterministic signals[J]. Proceedings of the IEEE, 1967, 55(4):523-531.), algorithms based on cyclostationarity characteristics (Sadeghi H, Azmi P. Performance analysis of linear cooperative cyclostationary spectrum sensing over Nakagami-m fading channels[J]. IEEE Transactions on Vehicular Technology, 2014, 63(9):4748-4756.), algorithms based on likelihood hypothesis (Soltanmohammadi E, Orooji M, Naraghi-Pour M. Spectrum sensing over MIMO channels using generalized likelihood ratio tests[J]. IEEE Signal Processing Letters, 2013, 20(5):439-442.), eigenvalue-based algorithms (Zeng Y, Liang YC. Eigenvalue-based spectrum sensing algorithms for cognitive radio[J]. IEEE Transactions on Communications, 2009, 57(6):1784-1793.).

[0004] In recent years, spectrum sensing based on deep learning has attracted much attention. et al. proposed a deep collaborative spectrum sensing scheme based on convolutional neural networks. This scheme adopts the strategy of autonomously learning to combine individual sensing results from training samples and considers the spatial and spectral correlations of channels. Subekti A et al. designed a spectrum sensing scheme using a deep autoencoder network and SVM (Subekti A, Pardede H, Sustika R and Suyoto. Spectrum sensing for cognitive radio using deep autoencoder neural network and SVM[C]. International Conference on Radar, Antenna, Microwave, Electronics, and Telecommunications(ICRAMET), Serpong, Indonesia, 2018:81-85.). Xie J et al. used CNN and long short-term memory networks to extract the spatial and temporal features of data and proposed a CNN-LSTM signal detector (Xie J, Fang J, Liu C, et al. Deep learning-based spectrum sensing in cognitive radio: a CNN-LSTM approach[J]. IEEE Communications Letters, 2020, 24(10):2196-2200.). Chen Z et al. proposed a blind spectrum sensing method based on short-time Fourier transform and deep learning (Chen Z, Xu Y Q, Wang H, et al. Deep STFT-CNN for spectrum sensing in cognitive radio[J]. IEEE Communications Letters, 2021, 25(3):864-868.). Liu R et al. established a detection framework based on CNN and extracted features based on the data-driven method of covariance matrix (Liu R, Ma Y, Zhang X, et al. Deep learning-based spectrum sensing in space-air-ground integrated networks[J]. Journal of Communications and Information Networks, 2021, 6(1):82-90.).Liu C et al. proposed a convolutional neural network spectrum sensing algorithm based on the covariance matrix (Liu C, Wang J, Liu X, et al. Deep CM-CNN for spectrum sensing in cognitive radio[J]. IEEE Journal on Selected Areas in Communications, 2019, 37(10): 2306-2321.). Xie J et al. proposed a spectrum sensing method based on unsupervised deep learning (Xie J, Fang J, Liu C, et al. Unsupervised deep spectrum sensing: A variational auto-encoder based approach[J]. IEEE Transactions on Vehicular Technology, 2020, 69(5): 5307-5319.).

[0005] The above spectrum sensing methods only consider the environmental noise as additive white Gaussian noise. However, in many practical applications, there are various non-Gaussian interferences, and these non-Gaussian interferences are usually modeled by the alpha-stable distribution. Since impulse interference often does not have a finite second moment, the detection performance of existing traditional methods degrades severely. In summary, the problems and defects of the existing technology are as follows:

[0006] (1) Traditional spectrum sensing methods only consider the influence of Gaussian noise or non-Gaussian interference, and few literatures mention spectrum sensing under Gaussian noise and non-Gaussian interference.

[0007] (2) The spectrum sensing method based on the deep convolutional network has the defect that the convolutional network cannot understand the global information of features. Summary of the Invention

[0008] To overcome the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide an intelligent spectrum sensing method, system, device and medium under non-Gaussian interference in a cognitive radio network. The method uses a non-linear preprocessor to process the observed signal and construct a generalized covariance matrix; uses a Transformer network to extract features from the constructed generalized covariance matrix and construct a feature vector based on the network output; uses the feature vectors under different hypothesis tests to construct a detection statistic and a detection threshold, and realizes spectrum sensing under non-Gaussian interference by comparing the detection statistic and the detection threshold, providing technical support for the cognitive radio network and effectively improving the anti-interference ability of the intelligent wireless system.

[0009] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0010] An intelligent spectrum sensing method under non-Gaussian interference in a cognitive radio network, which uses a non-linear preprocessor to process the observed signal and constructs a generalized covariance matrix; uses a Transformer network to extract features from the constructed generalized covariance matrix and constructs a feature vector based on the network output; uses the feature vectors under different hypothesis tests to construct a detection statistic and a detection threshold, and realizes spectrum sensing under non-Gaussian interference by comparing the detection statistic and the detection threshold.

[0011] An intelligent spectrum sensing method under non-Gaussian interference in a cognitive radio network, the specific steps of which include:

[0012] Step 1, use a non-linear preprocessor to process the observed signal r(n) and construct a generalized covariance matrix C r and a generalized time-frequency covariance matrix D r , and splice and combine the matrix C r and the matrix D r into a joint covariance matrix G r ;

[0013] Step 2, use a Transformer network to extract the deep features of the joint covariance matrix G r and construct a feature vector based on the network output

[0014] Step 3, use the feature vectors under different hypothesis tests to construct a detection statistic T vit and a detection threshold ψ, and realize spectrum sensing under non-Gaussian interference by comparing the detection statistic T vit and the detection threshold ψ.

[0015] Furthermore, the specific method of Step 1 is:

[0016] Based on a cognitive multi-antenna system, assume that the number of primary users and the number of antennas configured by cognitive users are P and M respectively, and the signal received by the m-th antenna is expressed as:

[0017]

[0018] where, I m (n) represents non-Gaussian interference, v m (n) represents additive Gaussian noise, h m,p represents the fading channel between the m-th antenna of the secondary user and the p-th primary user, and the observed signal is shown in the following matrix form:

[0019] r(n) = Hs(n) + Ι(n) + v(n);

[0020] where, H represents the fading channel matrix, s(n) = [s1(n),..., sP (n)] T denotes the transmitted signal matrix, I(n) = [I1(n),..., I M (n)] T denotes the non-Gaussian interference matrix, v(n) = [v1(n),..., v M (n)] T denotes the additive Gaussian noise matrix;

[0021] The non-Gaussian is characterized by the alpha-stable distribution, and its characteristic function expression is:

[0022]

[0023] In the formula,

[0024]

[0025] where α is called the characteristic exponent, used to measure the thickness of the tail of the distribution function; γ is called the dispersion coefficient; β is called the symmetry parameter, a is called the location parameter, β = 0 indicates that the distribution is a symmetric alpha-stable distribution SαS; if a = 0 and γ = 1, then this stable distribution is called the standard alpha-stable distribution;

[0026] The signal-to-interference-plus-noise ratio is defined as:

[0027]

[0028] where r(n) = [r1(n),..., r M (n)] T , v(n) = [v1(n),..., v M (n)] T .

[0029] Calculate the generalized covariance matrix

[0030]

[0031]

[0032]

[0033]

[0034] where, τ0 is usually a constant, τ0 = 1, is the median of the absolute value of the m-th row received signal.

[0035] Then, calculate the generalized time-frequency covariance matrix

[0036]

[0037] Among them, represents P rr the matrix element of P(n,f)

[0038]

[0039]

[0040] Among them, represents the Fourier transform, Q(n,τ) represents the time-delay kernel function, and where g(n,f) is a smoothing window of variables n and f.

[0041] Finally, the matrix C r and the matrix D r are spliced and combined into the joint covariance matrix G r

[0042] G r =[C r D r

[0043] Furthermore, the specific method of step two is as follows:

[0044] First, normalize the real part and the imaginary part of the covariance matrix respectively and form a two-channel matrix, which is used as the input of the Vision Transformer network;

[0045] Then, the input matrix is sliced into non-overlapping fixed-scale Patches, and each Patch is stretched into a one-dimensional vector. The stretched Patch sequence is subjected to a linear projection transformation layer to obtain the vector (token) corresponding to each Patch. Then, an optional learnable Class token is introduced, and its feature in the last layer of the encoder is used as the final feature for the classification layer. The learnable Class token is inserted at the beginning position of the vector sequence corresponding to each Patch to form the Input Embedding, and then the above sequence is added to the learnable position encoding to obtain the feature map; the obtained feature map is input into the encoder module to calculate the global attention and extract features. After passing through the repeated Encoder layers, the input of each layer of the encoder is the output of the previous layer of the encoder, and the feature map output by the last layer of the encoder is obtained;

[0046] Finally, the 0th position (Class token position) of the output sequence of the last encoder is extracted using the slicing operation to obtain the output feature map of the Class token. The obtained feature map is processed using a multi-layer perceptron, and the final feature vector is obtained through the Softmax function

[0047] Further, the specific method of Step 3 is as follows:

[0048] Construct a detection statistic T using the eigenvectors under different hypothesis tests vit and a detection threshold ψ. By comparing the detection statistic T vit and the detection threshold ψ, spectrum sensing under non-Gaussian interference is achieved:

[0049] Express the detection statistic as

[0050]

[0051] Express the detection threshold as

[0052]

[0053] If T vit > ψ, it is determined that the primary user signal exists; otherwise, it is determined that the primary user signal does not exist.

[0054] Further, based on the above intelligent spectrum sensing system under non-Gaussian interference in the cognitive radio network, it includes:

[0055] Matrix construction module 1, which is used to process the observed signal r(n) using a non-linear preprocessor to construct a generalized covariance matrix C r and a generalized time-frequency covariance matrix D r , and splice and combine the matrix C r and the matrix D r into a joint covariance matrix G r ;

[0056] Feature extraction module 2, which is used to extract the deep features of the joint covariance matrix G r using a Transformer network and construct an eigenvector based on the network output

[0057] Spectrum sensing module 3, which is used to construct a detection statistic T using the eigenvectors under different hypothesis tests vit and a detection threshold ψ. By comparing the detection statistic T vit and the detection threshold ψ, spectrum sensing under non-Gaussian interference is achieved. Further, an intelligent spectrum sensing device under non-Gaussian interference in a cognitive radio network is characterized by including:

[0058] A memory for storing computer programs;

[0059] A processor for implementing the intelligent spectrum sensing method under non-Gaussian interference in the cognitive radio network described in any one of Steps 1 to 3 when executing the computer program.

[0060] Furthermore, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can perform spectrum intelligent sensing under non-Gaussian interference in a cognitive radio network.

[0061] The beneficial effects of the present invention are as follows:

[0062] 1. Construct a feature matrix adapted to non-Gaussian pulse interference, and use the Transformer network to mine the global information in the deep layer of features. Then, use the binary hypothesis testing model to realize spectrum sensing under non-Gaussian interference, which can not only expand the applicable scenario range of intelligent wireless systems, but also effectively improve the anti-interference ability of intelligent wireless systems.

[0063] 2. The intelligent spectrum sensing method under non-Gaussian interference provided by the present invention uses the joint covariance matrix to characterize the signal and adopts the Transformer architecture based on the attention mechanism to mine the deep features of the signal, which can effectively realize spectrum intelligent sensing under Gaussian white noise and non-Gaussian pulse interference conditions, improve the spectrum sensing performance, and thus provide technical support for cognitive radio networks; the present invention has strong adaptability to alpha-stable distribution noise, and as the signal-to-noise ratio increases, the sensing performance gradually improves. Description of the Drawings

[0064] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required to be used in the embodiments of the present invention. Obviously, the following described drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0065] Figure 1 It is a flowchart of the intelligent spectrum sensing method under non-Gaussian interference provided by the embodiment of the present invention.

[0066] Figure 2 It is a schematic diagram of the intelligent spectrum sensing performance under non-Gaussian interference provided by the embodiment of the present invention. Detailed Embodiments

[0067] In order to make the objectives, technical solutions and advantages of the present invention clearer, the following further details the present invention with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0068] Aiming at the problems existing in the prior art, the present invention provides an intelligent spectrum sensing method and system under non-Gaussian interference in a cognitive radio network, and the following makes a detailed description of the present invention with reference to the drawings.

[0069] As Figure 1As shown in the figure, an intelligent spectrum sensing method under non-Gaussian interference in a cognitive radio network provided by an embodiment of the present invention includes the following steps:

[0070] S101. Process the observation signal r(n) using a non-linear pre-processor to construct a generalized covariance matrix C r and a generalized time-frequency covariance matrix D r , and splice and combine the matrix C r and the matrix D r into a joint covariance matrix G r ;

[0071] S102. Use a Transformer network to extract the deep features of the joint covariance matrix G r and construct a feature vector based on the network output

[0072] S103. Use the feature vectors under different hypothesis tests to construct a detection statistic T vit and a detection threshold ψ, and realize spectrum sensing under non-Gaussian interference by comparing the detection statistic T vit with the detection threshold ψ.

[0073] An intelligent spectrum sensing system under non-Gaussian interference in a cognitive radio network provided by an embodiment of the present invention includes:

[0074] A matrix construction module 1, configured to process the observation signal r(n) using a non-linear pre-processor to construct a generalized covariance matrix C r and a generalized time-frequency covariance matrix D r , and splice and combine the matrix C r and the matrix D r into a joint covariance matrix G r ;

[0075] A feature extraction module 2, configured to use a Transformer network to extract the deep features of the joint covariance matrix G r and construct a feature vector based on the network output

[0076] A spectrum sensing module 3, configured to use the feature vectors under different hypothesis tests to construct a detection statistic T vit and a detection threshold ψ, and realize spectrum sensing under non-Gaussian interference by comparing the detection statistic T vit with the detection threshold ψ.

[0077] The present invention will be further described below in conjunction with embodiments.

[0078] Embodiment 1

[0079] An intelligent spectrum sensing method under non-Gaussian interference in a cognitive radio network provided by an embodiment of the present invention includes the following steps:

[0080] S101, using a non-linear preprocessor to process the observed signal r(n) to construct a generalized covariance matrix C r and a generalized time-frequency covariance matrix D r , and splicing and combining the matrix C r and the matrix D r into a joint covariance matrix G r .

[0081] Based on a multi-antenna cognitive system, assuming that the number of primary users and the number of antennas configured by cognitive users are P and M respectively, the signal received by the m-th antenna is expressed as:

[0082]

[0083] where, I m (n) represents non-Gaussian interference, v m (n) represents additive Gaussian noise, h m,p represents the fading channel between the m-th antenna of the secondary user and the p-th primary user, and the observed signal is shown in the following matrix form:

[0084] r(n) = Hs(n) + Ι(n) + v(n);

[0085] where, H represents the fading channel matrix, s(n) = [s1(n),..., s P (n)] T represents the transmitted signal matrix, I(n) = [I1(n),..., I M (n)] T represents the non-Gaussian interference matrix, v(n) = [v1(n),..., v M (n)] T represents the additive Gaussian noise matrix;

[0086] The non-Gaussian is characterized by the alpha-stable distribution, and its characteristic function expression is:

[0087]

[0088] In the formula,

[0089]

[0090] where, α is called the characteristic exponent, used to measure the thickness of the tail of the distribution function; γ is called the dispersion coefficient; β is called the symmetry parameter, a is called the location parameter, β = 0 indicates that the distribution is a symmetric alpha-stable distribution SαS; if a = 0, γ = 1, then this stable distribution is called a standard alpha-stable distribution;

[0091] Define the signal-to-interference-plus-noise ratio as:

[0092]

[0093] where r(n) = [r1(n),..., r M (n)] T , v(n) = [v1(n),..., v M (n)] T .

[0094] Use a non-linear pre-processor to process the observation signal r(n) to construct the generalized covariance matrix C r and the generalized time-frequency covariance matrix D r , and splice and combine the two into the joint covariance matrix G r ; First, calculate the generalized covariance matrix

[0095]

[0096]

[0097]

[0098]

[0099] where τ0 is usually a constant, τ0 = 1, is the median of the absolute value of the m-th row received signal.

[0100] Then, calculate the generalized time-frequency covariance matrix

[0101]

[0102] where denotes the matrix element of P rr (n,f)

[0103]

[0104]

[0105] where denotes the Fourier transform, Q(n,τ) denotes the time-delay kernel function, and where g(n,f) is a smoothing window of variables n and f.

[0106] Finally, splice and combine the matrix C r and the matrix D r into the joint covariance matrix G r

[0107] G r = [C r D r

[0108] S102. Extract the deep features of the joint covariance matrix G using the Transformer network, and construct the feature vector based on the network output r and construct the feature vector based on the network output

[0109] First, normalize the real and imaginary parts of the joint covariance matrix respectively and form a two-channel matrix, which is used as the input of the Vision Transformer network

[0110] Then, the input matrix is sliced into non-overlapping patches of a fixed scale, and each patch is stretched into a one-dimensional vector. The stretched patch sequence is passed through a linear projection transformation layer to obtain the vector (token) corresponding to each patch. Then, an optional learnable Class token is introduced, and its feature at the last layer of the encoder is used as the final feature for the classification layer. The learnable Class token is inserted at the beginning of the vector sequence corresponding to each patch to form the Input Embedding. Then, the above sequence is added to the learnable position encoding to obtain the feature map. The obtained feature map is input into the encoder module to calculate the global attention and extract features. After passing through the repeated Encoder layers, the input of each layer of the encoder is the output of the previous layer of the encoder, and the feature map output by the last layer of the encoder is obtained

[0111] Finally, the 0th position (Class token position) of the output sequence of the last encoder is extracted using slicing operation to obtain the output feature map of the Class token. The feature map output by the encoder is processed using a multi-layer perceptron, and the final feature vector is obtained through the Softmax function

[0112] S103. Construct the detection statistic T using the feature vectors under different hypothesis tests vit and the detection threshold ψ, and realize the spectrum sensing under non-Gaussian interference by comparing the detection statistic T vit and the detection threshold ψ:

[0113] The detection statistic is expressed as

[0114]

[0115] The detection threshold is expressed as

[0116] ​

[0117] If T vit > ψ, it is determined that the primary user signal exists; otherwise, it is determined that the primary user signal does not exist.

[0118] The technical effects of the present invention will be described in detail below in combination with simulations.

[0119] In order to evaluate the performance of the present invention, simulation verification is carried out. The primary user signal is a MIMO signal, its baseband modulation method is BPSK, the noise is Gaussian white noise, and the non-Gaussian pulse interference is generated by a symmetric stable distribution. 1000 pairs of data are generated for each signal-to-noise ratio, and the normalized generalized covariance matrix is calculated as the input data of the network. In addition, pure noise and interference data are generated, and the data processing method is the same as above. Finally, a test and training data set with a total sample number of 40,000 is generated. The loss function for training the network model uses the cross-entropy function. The training uses the adaptive moment estimation optimization algorithm, and the learning rate is set to 0.001. A Dropout layer that eliminates some parameters is added during the training process to prevent overfitting in network training. The detection performance of the method proposed in the present invention at different signal-to-noise ratios is as Figure 2 shown. It can be seen from Figure 2 that the method of the present invention has a strong adaptability to alpha-stable distribution noise, and as the signal-to-noise ratio increases, the detection performance of the proposed method gradually improves.

[0120] As described above, the above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention shall be covered by the protection scope of the present invention.

Claims

1. An intelligent spectrum sensing method under non-Gaussian interference in a cognitive radio network, characterized in that a non-linear pre-processor is used to process the observed signal, a generalized covariance matrix and a generalized time-frequency covariance matrix are constructed, and the two are spliced and combined into a joint covariance matrix; a Transformer network is used to extract features from the constructed joint covariance matrix, and a feature vector based on the network output is constructed; feature vectors under different hypothesis tests are used to construct a detection statistic and a detection threshold, and spectrum sensing under non-Gaussian interference is realized by comparing the detection statistic and the detection threshold. The specific steps of the intelligent spectrum sensing method are as follows: Step 1: Process the observed signal r(n) using a non-linear pre-processor to construct a generalized covariance matrix C r and a generalized time-frequency covariance matrix D r , and concatenate matrix C r and matrix D r to form a joint covariance matrix G r : Based on a cognitive multi-antenna system, let the number of primary users and the number of antennas configured by cognitive users be P and M respectively, and the signal received by the m-th antenna is expressed as: Among them, I m (n) represents non-Gaussian interference, v m (n) represents additive Gaussian noise, h m,p represents the fading channel between the m-th antenna of the secondary user and the p-th primary user. The observed signal is shown in the following matrix form: r(n) = Hs(n) + Ι(n) + v(n); Among them, H represents the fading channel matrix, s(n) = [s1(n),..., s P (n)] T represents the transmitted signal matrix, I(n) = [I1(n),..., I M (n)] T represents the non-Gaussian interference matrix, v(n) = [v1(n),..., v M (n)] T represents the additive Gaussian noise matrix; The non-Gaussian is characterized by the alpha-stable distribution, and its characteristic function expression is: In the formula, where α is called the characteristic exponent, used to measure the thickness of the tail of the distribution function; γ is called the dispersion coefficient; β is called the symmetry parameter, a is called the location parameter, and β = 0 indicates that the distribution is a symmetric alpha-stable distribution SαS; if a = 0 and γ = 1, then this stable distribution is called the standard alpha-stable distribution. The signal-to-interference ratio is defined as: Among them, r(n) = [r1(n), …, r M (n)] T , v(n) = [v1(n),..., v M (n)] T ; Calculate the generalized covariance matrix Among them, τ0 is usually a constant, τ0 = 1, is the median of the absolute value of the received signal in the m-th row; Then, calculate the generalized time-frequency covariance matrix Among them, represents P rr the matrix element of (n,f) Among them, denotes the Fourier transform, Q(n,τ) denotes the time-delay kernel function, and where g(n,f) is a smoothing window for variables n and f; Finally, matrix C r and matrix D r are concatenated and combined into a joint covariance matrix G r G r = [C r D r ​ Step 2: Use the Transformer network to extract the deep features of the joint covariance matrix G r and construct a feature vector based on the network output First, the real and imaginary parts of the joint covariance matrix are normalized respectively and combined into a two-channel matrix, which is used as the input of the VisionTransformer network; Then, the input matrix is sliced into non-overlapping fixed-scale Patches, and each Patch is stretched into a one-dimensional vector. The stretched Patch sequence is passed through a linear projection transformation layer to obtain a vector (token) corresponding to each Patch; then, an optional learnable Class token is introduced, and its feature in the last layer of the encoder is used as the final feature for the classification layer; the learnable Class token is inserted at the beginning of the vector sequence corresponding to each Patch to form an Input Embedding, and then the above sequence is added to the learnable position encoding to obtain a feature map; the feature map obtained above is input into the encoder module to calculate global attention and extract features. After repeated Encoder layers, the input of each layer of the encoder is the output of the previous layer of the encoder, and the feature map output by the last layer of the encoder is obtained. Finally, extract the 0th position (Class token position) of the last encoder output sequence using slicing operation to obtain the output feature map of the Class token; process the feature map output by the encoder using a multi-layer perceptron and obtain the final feature vector through the Softmax function Step 3: Construct a detection statistic T using the eigenvectors under different hypothesis tests vit and a detection threshold ψ, and achieve spectrum sensing under non-Gaussian interference by comparing the detection statistic T vit with the detection threshold ψ.

2. The intelligent spectrum sensing method under non-Gaussian interference in a cognitive radio network according to claim 1, wherein The specific method of step three is: Construct the detection statistic T using the eigenvectors under different hypothesis tests vit and the detection threshold ψ, and realize spectrum sensing under non-Gaussian interference by comparing the detection statistic T vit and the detection threshold ψ: The detection statistic is expressed as The detection threshold is expressed as If T vit > ψ, it is determined that the primary user signal exists; otherwise, it is determined that the primary user signal does not exist.

3. An intelligent spectrum sensing system under non-Gaussian interference in a cognitive radio network based on the intelligent spectrum sensing method according to any one of claims 1 or 2, characterized in that, including: The matrix construction module 1 is used to process the observed signal r(n) by means of a non-linear preprocessor to construct a generalized covariance matrix C r and a generalized time-frequency covariance matrix D r , and combine the matrix C r and the matrix D r by splicing to form a joint covariance matrix G r ; The feature extraction module 2 is used to extract the deep features of the joint covariance matrix G by using the Transformer network and construct a feature vector based on the network output r ​ The spectrum sensing module 3 is used to construct a detection statistic T using the eigenvectors under different hypothesis tests vit and a detection threshold ψ, and realizes spectrum sensing under non-Gaussian interference by comparing the detection statistic T vit with the detection threshold ψ.

4. An intelligent spectrum sensing device under non-Gaussian interference in a cognitive radio network, characterized in that, including: a memory for storing a computer program; a processor for implementing the intelligent spectrum sensing method under non-Gaussian interference in the cognitive radio network according to any one of claims 1 to 2 when executing the computer program.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it can perform intelligent spectrum sensing under non-Gaussian interference in the cognitive radio network according to any one of claims 1 to 2.

Images